Benchmarks

Benchmarking the robustness of state-of-the-art methods in all seven scenarios of the nuScenes-R and Waymo-R.

Method Modality Performance LiDAR Camera
P_c mP_R R Stuck FOV Object Stuck Missing Occlusion Calib
nuScenes-R (mAP / NDS)
CenterPoint L 56.8/65.0 23.4/46.3 0.41/0.71 26.1/47.5 15.6/43.0 28.4/48.5 - - - -
DETR3D C 34.9/43.4 17.6/31.5 0.50/0.73 - - - 17.3/32.3 14.5/29.9 14.3/29.0 24.2/35.0
PointAugmenting LC 46.9/55.6 33.7/48.3 0.72/0.87 25.3/43.5 13.3/37.7 21.3/39.4 42.1/52.8 37.0/49.8 40.7/52.2 43.6/53.8
MVX-Net LC 61.0/66.1 38.4/53.6 0.63/0.81 35.2/51.4 17.6/43.1 34.0/51.1 48.3/58.8 32.7/50.6 45.5/57.6 50.8/59.9
TransFusion LC 66.9/70.9 52.8/63.1 0.79/0.89 33.4/52.3 20.3/45.8 34.6/53.6 65.9/70.2 64.9/69.7 65.5/70.0 66.5/70.7
Waymo-R(L2 mAP / L2 mAPH)
CenterPoint L 66.0/63.4 30.6/29.4 0.46/0.46 29.5/28.3 30.3/29.1 32.1/30.9 - - - -
DETR3D C 16.2/15.7 10.1/9.8 0.62/0.62 - - - 13.0/12.6 8.4/8.2 10.9/10.5 8.0/7.8
PointAugmenting LC 52.5/50.7 39.6/38.3 0.75/0.76 24.7/23.9 24.3/23.4 26.2/25.3 51.7/50.0 50.4/48.6 50.3/48.6 49.8/48.1
MVX-Net LC 59.7/54.1 44.3/40.1 0.74/0.74 27.5/24.9 28.8/25.6 28.7/26.0 58.2/52.7 55.9/50.5 56.4/51.1 54.9/49.6
TransFusion LC 66.7/64.1 51.2/49.1 0.77/0.77 30.2/29.0 30.2/29.0 32.7/31.3 66.5/63.9 66.1/63.5 66.2/63.6 66.3/63.7